Evaluating Search Cost Models: Estimation and Prediction

The classic search models assume that consumers adhere to a particular method of search (sequential or non-sequential) and that they know the true price distribution. In this paper, we evaluate how well the search cost estimates from classic models predict search outcomes - the amount of search and purchase prices - when these assumptions are violated. To this end, we conduct an online experiment in which we vary searchers' information about the price distribution of a homogeneous good. For each treatment, we (i) estimate search costs, (ii) fit each model to the estimated search cost distribution to obtain in- and out-of-sample predictions about outcomes, and (iii) compare predicted and realized outcomes. We find that the prediction performance of each model is largely robust to violations of the informational assumption. Further, the prediction performance of the sequential and non-sequential search model are similar, despite the fact that the search environment strongly favors sequential search.